How to handle a signal.SIGINT on a Windows OS machine?
I am trying the code pasted below on Windows, but instead of handling signal, it is killing the process. However, the same code is working in Ubuntu.
import os, sys
import time
import signal
def func(signum, frame):
print 'You raised a SigInt! Signal handler called with signal', signum
signal.signal(signal.SIGINT, func)
while True:
print "Running...",os.getpid()
time.sleep(2)
os.kill(os.getpid(),signal.SIGINT)
Solution 1:
Python's os.kill wraps two unrelated APIs on Windows. It calls GenerateConsoleCtrlEvent when the sig parameter is CTRL_C_EVENT or CTRL_BREAK_EVENT. In this case the pid parameter is a process group ID. If the latter call fails, and for all other sig values, it calls OpenProcess and then TerminateProcess. In this case the pid parameter is a process ID, and the sig value is passed as the exit code. Terminating a Windows process is akin to sending SIGKILL to a POSIX process. Generally this should be avoided since it doesn't allow the process to exit cleanly.
Note that the docs for os.kill mistakenly claim that "kill() additionally takes process handles to be killed", which was never true. It calls OpenProcess to get a process handle.
The decision to use WinAPI CTRL_C_EVENT and CTRL_BREAK_EVENT, instead of SIGINT and SIGBREAK, is unfortunate for cross-platform code. It's also not defined what GenerateConsoleCtrlEvent does when passed a process ID that's not a process group ID. Using this function in an API that takes a process ID is dubious at best, and potentially very wrong.
For your particular needs you can write an adapter function that makes os.kill a bit more friendly for cross-platform code. For example:
import os
import sys
import time
import signal
if sys.platform != 'win32':
kill = os.kill
sleep = time.sleep
else:
# adapt the conflated API on Windows.
import threading
sigmap = {signal.SIGINT: signal.CTRL_C_EVENT,
signal.SIGBREAK: signal.CTRL_BREAK_EVENT}
def kill(pid, signum):
if signum in sigmap and pid == os.getpid():
# we don't know if the current process is a
# process group leader, so just broadcast
# to all processes attached to this console.
pid = 0
thread = threading.current_thread()
handler = signal.getsignal(signum)
# work around the synchronization problem when calling
# kill from the main thread.
if (signum in sigmap and
thread.name == 'MainThread' and
callable(handler) and
pid == 0):
event = threading.Event()
def handler_set_event(signum, frame):
event.set()
return handler(signum, frame)
signal.signal(signum, handler_set_event)
try:
os.kill(pid, sigmap[signum])
# busy wait because we can't block in the main
# thread, else the signal handler can't execute.
while not event.is_set():
pass
finally:
signal.signal(signum, handler)
else:
os.kill(pid, sigmap.get(signum, signum))
if sys.version_info[0] > 2:
sleep = time.sleep
else:
import errno
# If the signal handler doesn't raise an exception,
# time.sleep in Python 2 raises an EINTR IOError, but
# Python 3 just resumes the sleep.
def sleep(interval):
'''sleep that ignores EINTR in 2.x on Windows'''
while True:
try:
t = time.time()
time.sleep(interval)
except IOError as e:
if e.errno != errno.EINTR:
raise
interval -= time.time() - t
if interval <= 0:
break
def func(signum, frame):
# note: don't print in a signal handler.
global g_sigint
g_sigint = True
#raise KeyboardInterrupt
signal.signal(signal.SIGINT, func)
g_kill = False
while True:
g_sigint = False
g_kill = not g_kill
print('Running [%d]' % os.getpid())
sleep(2)
if g_kill:
kill(os.getpid(), signal.SIGINT)
if g_sigint:
print('SIGINT')
else:
print('No SIGINT')
Discussion
Windows doesn't implement signals at the system level [*]. Microsoft's C runtime implements the six signals that are required by standard C: SIGINT, SIGABRT, SIGTERM, SIGSEGV, SIGILL, and SIGFPE.
SIGABRT and SIGTERM are implemented just for the current process. You can call the handler via C raise. For example (in Python 3.5):
>>> import signal, ctypes
>>> ucrtbase = ctypes.CDLL('ucrtbase')
>>> c_raise = ucrtbase['raise']
>>> foo = lambda *a: print('foo')
>>> signal.signal(signal.SIGTERM, foo)
<Handlers.SIG_DFL: 0>
>>> c_raise(signal.SIGTERM)
foo
0
SIGTERM is useless.
You also can't do much with SIGABRT using the signal module because the abort function kills the process once the handler returns, which happens immediately when using the signal module's internal handler (it trips a flag for the registered Python callable to be called in the main thread). For Python 3 you can instead use the faulthandler module. Or call the CRT's signal function via ctypes to set a ctypes callback as the handler.
The CRT implements SIGSEGV, SIGILL, and SIGFPE by setting a Windows structured exception handler for the corresponding Windows exceptions:
STATUS_ACCESS_VIOLATION SIGSEGV
STATUS_ILLEGAL_INSTRUCTION SIGILL
STATUS_PRIVILEGED_INSTRUCTION SIGILL
STATUS_FLOAT_DENORMAL_OPERAND SIGFPE
STATUS_FLOAT_DIVIDE_BY_ZERO SIGFPE
STATUS_FLOAT_INEXACT_RESULT SIGFPE
STATUS_FLOAT_INVALID_OPERATION SIGFPE
STATUS_FLOAT_OVERFLOW SIGFPE
STATUS_FLOAT_STACK_CHECK SIGFPE
STATUS_FLOAT_UNDERFLOW SIGFPE
STATUS_FLOAT_MULTIPLE_FAULTS SIGFPE
STATUS_FLOAT_MULTIPLE_TRAPS SIGFPE
The CRT's implementation of these signals is incompatible with Python's signal handling. The exception filter calls the registered handler and then returns EXCEPTION_CONTINUE_EXECUTION. However, Python's handler only trips a flag for the interpreter to call the registered callable sometime later in the main thread. Thus the errant code that triggered the exception will continue to trigger in an endless loop. In Python 3 you can use the faulthandler module for these exception-based signals.
That leaves SIGINT, to which Windows adds the non-standard SIGBREAK. Both console and non-console processes can raise these signals, but only a console process can receive them from another process. The CRT implements this by registering a console control event handler via SetConsoleCtrlHandler.
The console sends a control event by creating a new thread in an attached process that begins executing at CtrlRoutine in kernel32.dll or kernelbase.dll (undocumented). That the handler doesn't execute on the main thread can lead to synchronization problems (e.g. in the REPL or with input). Also, a control event won't interrupt the main thread if it's blocked while waiting on a synchronization object or waiting for synchronous I/O to complete. Care needs to be taken to avoid blocking in the main thread if it should be interruptible by SIGINT. Python 3 attempts to work around this by using a Windows event object, which can also be used in waits that should be interruptible by SIGINT.
When the console sends the process a CTRL_C_EVENT or CTRL_BREAK_EVENT, the CRT's handler calls the registered SIGINT or SIGBREAK handler, respectively. The SIGBREAK handler is also called for the CTRL_CLOSE_EVENT that the console sends when its window is closed. Python defaults to handling SIGINT by rasing a KeyboardInterrupt in the main thread. However, SIGBREAK is initially the default CTRL_BREAK_EVENT handler, which calls ExitProcess(STATUS_CONTROL_C_EXIT).
You can send a control event to all processes attached to the current console via GenerateConsoleCtrlEvent. This can target a subset of processes that belong to a process group, or target group 0 to send the event to all processes attached to the current console.
Process groups aren't a well-documented aspect of the Windows API. There's no public API to query the group of a process, but every process in a Windows session belongs to a process group, even if it's just the wininit.exe group (services session) or winlogon.exe group (interactive session). A new group is created by passing the creation flag CREATE_NEW_PROCESS_GROUP when creating a new process. The group ID is the process ID of the created process. To my knowledge, the console is the only system that uses the process group, and that's just for GenerateConsoleCtrlEvent.
What the console does when the target ID isn't a process group ID is undefined and should not be relied on. If both the process and its parent process are attached to the console, then sending it a control event basically acts like the target is group 0. If the parent process isn't attached to the current console, then GenerateConsoleCtrlEvent fails, and os.kill calls TerminateProcess. Weirdly, if you target the "System" process (PID 4) and its child process smss.exe (session manager), the call succeeds but nothing happens except that the target is mistakenly added to the list of attached processes (i.e. GetConsoleProcessList). It's probably because the parent process is the "Idle" process, which, since it's PID 0, is implicitly accepted as the broadcast PGID. The parent process rule also applies to non-console processes. Targeting a non-console child process does nothing -- except mistakenly corrupt the console process list by adding the unattached process. I hope it's clear that you should only send a control event to either group 0 or to a known process group that you created via CREATE_NEW_PROCESS_GROUP.
Don't rely on being able to send CTRL_C_EVENT to anything but group 0, since it's initially disabled in a new process group. It's not impossible to send this event to a new group, but the target process first has to enable CTRL_C_EVENT by calling SetConsoleCtrlHandler(NULL, FALSE).
CTRL_BREAK_EVENT is all you can depend on since it can't be disabled. Sending this event is a simple way to gracefully kill a child process that was started with CREATE_NEW_PROCESS_GROUP, assuming it has a Windows CTRL_BREAK_EVENT or C SIGBREAK handler. If not, the default handler will terminate the process, setting the exit code to STATUS_CONTROL_C_EXIT. For example:
>>> import os, signal, subprocess
>>> p = subprocess.Popen('python.exe',
... stdin=subprocess.PIPE,
... creationflags=subprocess.CREATE_NEW_PROCESS_GROUP)
>>> os.kill(p.pid, signal.CTRL_BREAK_EVENT)
>>> STATUS_CONTROL_C_EXIT = 0xC000013A
>>> p.wait() == STATUS_CONTROL_C_EXIT
True
Note that CTRL_BREAK_EVENT wasn't sent to the current process, because the example targets the process group of the child process (including all of its child processes that are attached to the console, and so on). If the example had used group 0, the current process would have been killed as well since I didn't define a SIGBREAK handler. Let's try that, but with a handler set:
>>> ctrl_break = lambda *a: print('^BREAK')
>>> signal.signal(signal.SIGBREAK, ctrl_break)
<Handlers.SIG_DFL: 0>
>>> os.kill(0, signal.CTRL_BREAK_EVENT)
^BREAK
[*]
Windows has asynchronous procedure calls (APC) to queue a target function to a thread. See the article Inside NT's Asynchronous Procedure Call for an in-depth analysis of Windows APCs, especially to clarify the role of kernel-mode APCs. You can queue a user-mode APC to a thread via QueueUserAPC. They also get queued by ReadFileEx and WriteFileEx for the I/O completion routine.
A user-mode APC executes when the thread enters an alertable wait (e.g. WaitForSingleObjectEx or SleepEx with bAlertable as TRUE). Kernel-mode APCs, on the other hand, get dispatched immediately (when the IRQL is below APC_LEVEL). They're typically used by the I/O manager to complete asynchronous I/O Request Packets in the context of the thread that issued the request (e.g. copying data from the IRP to a user-mode buffer). See Waits and APCs for a table that shows how APCs affect alertable and non-alertable waits. Note that kernel-mode APCs don't interrupt a wait, but instead are executed internally by the wait routine.
Windows could implement POSIX-like signals using APCs, but in practice it uses other means for the same ends. For example:
-
Structured Exception Handling, e.g.
__try,__except,__finally,__leave,RaiseException,AddVectoredExceptionHandler. Kernel Dispatcher Objects (i.e. Synchronization Objects), e.g.
SetEvent,SetWaitableTimer.Window Messages, e.g.
SendMessage(to a window procedure),PostMessage(to a thread's message queue to be dispatched to a window procedure),PostThreadMessage(to a thread's message queue),WM_CLOSE,WM_TIMER.
Window messages can be sent and posted to all threads that share the calling thread's desktop and that are at the same or lower integrity level. Sending a window message puts it in a system queue to call the window procedure when the thread calls PeekMessage or GetMessage. Posting a message adds it to the thread's message queue, which has a default quota of 10,000 messages. A thread with a message queue should have a message loop to process the queue via GetMessage and DispatchMessage. Threads in a console-only process typically do not have a message queue. However, the console host process, conhost.exe, obviously does. When the close button is clicked, or when the primary process of a console is killed via the task manager or taskkill.exe, a WM_CLOSE message is posted to the message queue of the console window's thread. The console in turns sends a CTRL_CLOSE_EVENT to all of its attached processes. If a process handles the event, it's given 5 seconds to exit gracefully before it's forcefully terminated.